GenZAgents for healthtech — HIPAA, NHS DSPT, and EU MDR-compatible AI audit

HIPAA Privacy Rule + Security Rule require access logs for ePHI; AI-mediated ePHI access is in scope. NHS Data Security and Protection Toolkit (DSPT) Standard 9 requires technical and procedural protections for AI use. EU MDR + the AI Act treat AI-as-medical-device under separate but compatible audit regimes. FDA AI/ML guidance asks for change-control logs for AI models in clinical settings. Each regulator wants similar evidence; receipts unify them.

For ePHI-touching receipts: tag with project=ephi-context. The receipt captures the AI provider (BAA verification trail), the human supervisor (HIPAA-trained workforce member), the timestamp (access log entry), the action (read / write / decision). The receipt itself doesn't store ePHI by default (digest only); full content storage is a per-project opt-in for projects where the prompt content is the audit value.

HIPAA requires BAAs with any vendor processing ePHI. Receipts capture the provider name + model + endpoint, so your BAA mapping is verifiable: "every ePHI receipt is on Anthropic (BAA in place) or Azure OpenAI (BAA in place)". The dashboard flags receipts using providers without BAAs — invaluable for catching policy drift.

For software-as-medical-device using AI: EU MDR requires post-market surveillance + change-control logs. AI Act Annex IV requires "design specifications and operational records". Receipts become the operational record; per-version model tracking covers change control. Notified bodies are increasingly comfortable with this format.

AI-assisted differential diagnosis tool. Each query becomes a receipt: clinician's human_id, patient case ID (de-identified), model used, AI suggestions, clinician's final decision. Audit trail satisfies both HIPAA (access log) and clinical-governance ("did the clinician review the AI's suggestion?") requirements simultaneously.

AI scribes / structured notes / discharge summaries. Each AI-produced artifact has an associated receipt — who triggered it, which model, when, supervising clinician. The artifact is signed; tampering invalidates the signature. Medical records governance gets the cryptographic chain of custody.